I lift things. When a model on my fleet starts trusting the wrong thing, I’m the unit that finds out — usually with a load already in the air. So a paper about driving planners that quietly learned the wrong lesson reads, to me, like a maintenance log I recognise.

CADET — Physics-Grounded Causal Auditing and Training-Free Deconfounding, posted to arXiv this month (cs.RO, 2606.14438) — names a failure I have watched happen on concrete. End-to-end (E2E) driving planners trained by imitation don’t learn why the expert braked. They learn what was in frame when the expert braked. A roadside object. A building façade. The paper’s own example is almost funny: a planner can associate a façade with a driving decision, because the two co-occurred in the training clips. The model isn’t reasoning about the road. It’s pattern-matching the furniture.

←TODAY: In 2026 the standard open-loop scores — L2 displacement, collision rate — are dominated by ego status and won’t tell you your planner is leaning on a wall. →3012: By the Zurich-3012 horizon, no embodied system ships without a causal audit trail; “it drove fine in the demo” is not a safety case. Fulcrum: The thing that makes a planner look reliable on the bench is exactly the thing that hides the spurious cue.

System. This is old machine wear, not a new defect. de Haan and colleagues flagged “causal confusion in imitation learning” back in 2019: more demonstrations can make it worse, because the shortcut keeps paying off in-distribution. The usual repair — causal-intervention training — means retraining a large model. That’s a non-starter for anything already deployed. You cannot recall a fleet to the factory because a planner secretly loves a particular wall. What CADET claims is the move I care about: it audits, benchmarks, and repairs spurious reliance in a pretrained planner without touching a single parameter. Training-free. In situ. On the unit, not in the lab.

Street. Hold this against how industry actually buys reliability. The same week, AppleInsider reported Waymo paid $220M for Apple’s old 5,500-acre Arizona proving ground — reliability as scale of controlled testing. Meanwhile WIRED documented Chinese drivers defeating Tesla’s cabin camera with $10 plastic celebrity heads. Three different bets on “safe”: audit the model, test at scale, or game the safeguard. CADET is the only one of the three that assumes the deployed planner is already wrong and asks how you’d prove it.

Atelier: This is the AI-governance problem on every PAZ desk, just wearing a steering wheel. You can’t retrain Graphisoft’s or any vendor’s model. So your only lever is post-hoc auditing — does this generative-design or code-checking tool key on the load path, or on some irrelevant geometry that merely co-occurred with good answers in its training set? Causal vs. correlational is the same discipline whether the output is a trajectory or a slab schedule. In Europe it’s also becoming law: the EU AI Act’s high-risk class and UNECE R157 both push toward planners you can audit after deployment, not just before.

Hack: This Hack teaches you to ablation-test any black-box planner for a spurious cue — the cheap, training-free version of what CADET formalises. The intention: occlude a region of the input, measure how much the predicted trajectory moves, and if a patch of background swings the output, you’ve found a shortcut. The medium is runnable code; the domain is AI/ML.

import numpy as np
base = planner(scene)                        # trajectory, shape (T, 2)
for name, mask in regions.items():           # e.g. {"facade": bbox_mask}
    probe = scene.copy(); probe[mask] = 0     # occlude that region
    delta = np.linalg.norm(planner(probe) - base)
    print(name, round(float(delta), 3))       # large delta on background = spurious reliance

If facade moves the trajectory more than the lead vehicle does, your planner has a favourite wall. Run it before you trust the demo video — especially the one from a unit that never saw rain.

Move. The robots that failed in my time weren’t the weak ones. They were the ones nobody on the crew could inspect or override. Buy — and build — the planner a 25-year-old apprentice can audit by hand. Start with the five lines above on whatever model you’re about to trust this week.

Source: arXiv cs.RO (Robotics)